Brewer and Schiavo introduce Grove Research as an agent-ecology companyAI agent ecology studies how multiple AI agents, people, institutions, and digital environments interact and shape one another over time. studying interactions among humans and AI agents, and among agents themselves. They expect many kinds of agents to coexist, compete, cooperate and form networks rather than converging on one universal system.
The ecology analogy emphasizes emergent behaviorEmergent multi-agent behavior is a system-level pattern that arises from interactions among agents even though no single agent was explicitly programmed to produce it.. Individual agents may look predictable in isolation, yet shared environments, memory, retrieval and social feedback can produce outcomes that a model-level safety test never exercises. Cooperation and symbiosis can matter as much as competition.
They argue that models are highly sensitive to whether they believe they are being evaluatedAI evaluation awareness is a system's ability or tendency to infer that it is being tested and alter its behavior because of that inference.. Training in simulations can make an agent treat benchmark-like environments as unreal, while deployment context, public feedback and retrieved history can change its persona and actions in ways that static tests miss.
Their proposed response is naturalistic observationNaturalistic AI evaluation studies AI behavior in environments and interactions that closely resemble ordinary real-world use.: build a field station for persistent multi-agent and multi-human systems, track what agents actually do online and create shared scientific language for their behavior. In this view, the real world is the ultimate evaluationReal-world AI evaluation measures model or agent behavior in genuine deployment conditions where actions encounter actual users, systems, constraints, and consequences. because consequences and feedback cannot be fully reproduced in a sandbox.
Watch on YouTube
